Искусственный General Int ...

Можете ли вы разработать игру, в которой люди всегда имеют больше преимуществ, чем ИИ, независимо от того, насколько он продвинут?

Любит | Нелюбов | Ответы | Вид | 1312


Шрираман Мадхаван Стэнфордская Статистика |   
@ | Обновлено Right Now


Anything involving humor/memes, I would confidently bet on humans to outperform.

A game with questions like “Is this sentence/picture a joke?”. Humor detection is by far, the most difficult task in natural language processing. Today’s AI systems built to detect humor are still almost clueless, and will be for the foreseeable future. Let’s look at different levels of humor to see why:

  1. Puns:
    Among the different kinds of humor, detecting puns is probably the simplest (and yet difficult), because it is a task in word sense disambiguation i.e., in the sentence “I’m a banker, but I lost interest”, the fact that the word interest is used in two different contexts at once (“lost interest”, banker-interest) is kinda detectable automatically.
  2. Sarcasm:
    A lot of work has been done on sarcasm detection, and state-of-the-art models are pretty good (85–90% accuracy on benchmark data sets). To get a sense of why that’s possible, consider the sentence “I loved the movie so much that I left the theater during the interval”. Here, there is a sentiment shift in the same sentence (“love movie” -> “left theater”), which seems to be an indicator of sarcasm. And that, and many more features like that, can be detected easily.
  3. Funny one-liners with no context:
    Most investigation in humor recognition has happened on funny one-liners - ones that are syntactically simple. Consider the one-liner “Infants don’t enjoy infancy half as much as adults enjoy adultery”. The humor relies on two things: Phonology (infancy-adultery) and the fact that the relationships between the words infant-infancy and adult-adultery are different. Both patterns are detectable automatically.
  4. Funny one-liners with context:
    This is where it gets tricky. For example, the one-liner “It was so cold last winter that I saw a lawyer with his hands in his own pockets” requires some context about lawyer stereotypes that aren’t self-evident from the one-liner. And the emphasis on “his own” which one would use while saying this particular joke, is completely missed out in text. AI systems perform terribly here.
  5. Memes:
    I genuinely believe that today’s internet memes are the pinnacle of human humor, although many may disagree. It is almost impossible for an AI system to detect humor in memes because a) It requires a lot of context and inferences, and b) There’s usually text and pictures. For example, consider this:

Source: UCLA Memes for Sick AF Tweens

Think about the context needed for an AI system to detect humor in that picture: a) The teacher is talking about a test; b) That person was confident that he didn’t get a 47; c) He realizes that he got a 47, and d) That’s the face people make when they’re shocked + embarrassed + sad.

(Я, вероятно, должен пояснить: система искусственного интеллекта, которая сегодня обучается на тысячах интернет-мемов, может определять, является ли данное изображение мемом или нет (поскольку мемы обычно имеют некоторые обнаруживаемые особенности: определенные изображения, определенный текст, появившийся в других мемах). и т. д.) Но обнаружение юмора в подобном мему изображении является невозможной задачей для современных систем ИИ. Например, замените то, что говорит учитель на изображении выше, на что-то другое, что не имеет смысла. Изображение не будет иметь смысла. быть смешным, но все же будет иметь мем-подобные функции, которые обнаружит система ИИ.)

Во многих отношениях юмор охватывает все, что делает человека человеком.

| |



Онлайн-курс
«Машинное обучение для людей»
Один из самых трендовых курсов в сфере Цифровой экономики.